AI Language Model Showdown: A Deep Dive into ChatGPT 3.5, ChatGPT 4, and Claude 2

In my role at LAIA Ventures, I have had the privilege of leading the AI working group at LAIA Observatory in a focused endeavor involving the integration of four APIs into our state-of-the-art ConTech (Construction Technologies) startup tracking tool. This ambitious project, known as ConTech Panel, places AI at its core, and our team has dedicated the past two months to its intensive development. Over the course of the past year, we have had the opportunity to collaborate with approximately 15 distinct specialized AI models, each dedicated to language modeling and text analysis.

Finally, we have chosen to integrate four commercial AI models that we consider to be better trained: Davinci-003, ChatGPT 3.5, ChatGPT 4, and Claude 2. The results obtained through systematic and extensive application for our specific use case have been very enlightening. They lead us to conclude that massive model training is fundamental to achieving more efficient, rich, and precise results.

?

Of all these models, ChatGPT 3.5 unequivocally provides the most balanced and "sensible" responses. While its successor, ChatGPT 4, is much more powerful and has excellent analytical capabilities, ChatGPT 3.5 delivers more natural, well-articulated, linguistically nuanced, rich, and precise arguments. Its supremacy in these areas is surprising and can only be attributed to its highly efficient training on massive conversational sources and fine-tuning based on extensive real-world usage, a level only achievable by OpenAI, the company behind it.

?

The foundations of ChatGPT 3.5 are rooted in the Davinci model, a brilliant AI engine that undoubtedly sparked the explosion of current conversational models. Davinci-003 is a model with a great wealth of language, sometimes barely contained by the engine, resulting in responses that can be overly expansive and lacking in order and judgment. Despite this "flaw," its results can be very interesting as a basis for work due to its multifaceted approach and the wellspring of instant ideas it provides. Its use is akin to a brainstorming exercise.

?

ChatGPT 4 is an advanced language model with greater capabilities than its predecessor. For our specific use cases, it has proven to be very efficient and satisfactory due to its analytical capability, ability to articulate ideas, and evaluate concepts. In this version of OpenAI's most advanced language model, considerable effort has been put into its development, and the advantages and differences compared to its predecessor are clear. However, its lack of training and lower commercial usage still weigh it down, resulting at times in less rich, less imaginative, more rigid, and less fluent responses compared to the almost "human" naturalness achieved by ChatGPT 3.5.

?

Claude 2 is one of the newer commercial players in the field of language model AI. Its developer, Anthropic, has a team that originated from OpenAI, and this resemblance is noticeable from the outset, given the structured nature of its responses, similar to those of ChatGPT 4. Its architecture and training models are quite similar, undoubtedly stemming from common engineering. However, there are significant differences that make it an interesting option. Based on our testing, its analytical capabilities have proven to be better in Claude 2, albeit not consistently. It is more precise and has greater evaluation and synthesis capabilities. This, however, is both its best feature and its worst flaw when the intended use requires linguistic richness and conceptual breadth. It suffers from a lack of training and commercial usage, which limits it in these aspects. Moreover, its significant ability to handle very broad contexts, while initially a great functionality, can be limiting in delivering results, often missing the essence or spirit of the texts it scrutinizes.

?

Claude 2 is perhaps the most advanced model, partly free from the common traits found in OpenAI's language models. However, it needs improvement through real-world usage. Currently, the competition between it and ChatGPT 4 is close, but this balance may change. The ability of one or the other to adapt to real-world uses will undoubtedly drive this change.

?

Nevertheless, ChatGPT 3.5 is currently the most refined strictly as a language model. It is a mistake to consider ChatGPT 4 a better option because it is a different model, more specialized and oriented toward a somewhat different, more analytical use, further removed from "human" mechanisms.